Cheaper, open models gain ground
Industry coverage argues that many customers are shifting away from frontier‑scale AI models toward cheaper, open‑weight or task‑specific systems that better fit cost and privacy constraints. The discussion frames model choice as an operational economics problem—firms weigh budget, governance and workflow fit rather than raw benchmark performance. (The Register)
Companies are increasingly choosing cheaper, open-weight or task-specific artificial intelligence models over the biggest closed systems when they put tools into production. (theregister.com) The shift is showing up in both pricing and model releases. OpenAI’s current API pricing lists GPT-5.4 at $2.50 per 1 million input tokens and $15 per 1 million output tokens, while vendors such as Google, IBM and Mistral are pushing smaller or open models as lower-cost deployment options. (openai.com (ibm.com) (mistral.ai) (blog.google)) An open-weight model is a system whose trained parameters can be downloaded and run by customers on their own hardware or in their own cloud account. That gives companies more control over where data goes, how long it is stored, and how tightly the model can be tuned to an internal workflow. (opensource.org) (mistral.ai) (deepmind.google) The performance gap has also narrowed enough that many buyers no longer need the largest model for every task. Google said last week that Gemma 4’s 31 billion-parameter model ranked as the No. 3 open model on the Arena leaderboard and beat models roughly 20 times its size on that measure. (blog.google) (arena.ai) That changes the buying math for routine work such as summarizing documents, extracting fields from forms, answering internal questions and generating code suggestions. IBM says its Granite line is built for “cost efficiency” and “enterprise workloads,” while Mistral markets open models that customers can customize and deploy with “full ownership” of data. (ibm.com 1) (ibm.com 2) (mistral.ai) The language around “open” still needs care. The Open Source Initiative published its Open Source AI Definition in October 2024, and many widely used “open-weight” models do not meet that stricter standard because access to training data or other components remains limited. (opensource.org) (techcrunch.com) Model choice is also becoming more granular inside one company. IBM says customers should pick the model that fits the use case, the brand and the budget, instead of defaulting to a single provider for every job. (ibm.com) That does not mean frontier closed models are disappearing. Anthropic still sells Claude Opus 4.6 for high-end workloads at $5 per million input tokens and $25 per million output tokens, and OpenAI continues to position GPT-5.4 as its most capable model for complex professional work. (anthropic.com) (openai.com) What is changing is where those expensive models sit in the stack. For many companies in 2026, the default is no longer “best benchmark at any cost,” but the cheapest model that can meet a security rule, stay inside a workflow and finish the job reliably. (theregister.com)
